Lorensius Forbin Dapa, Supatman
Visitor density monitoring in commercial areas directly affects safety and facility comfort management. This study develops a visitor density detection and classification system based on a pretrained YOLOv8s model, with a case study at Super Dazzle Gejayan, Yogyakarta. The YOLOv8s model was used without retraining, leveraging weights pretrained on the COCO dataset to detect human objects in real time. Detection results were then classified into three density categories: Quiet (≤5 people), Moderate (6–10 people), and Crowded (>10 people), using a threshold-based rule approach. Testing was conducted on 707 static images collected directly from the store environment under various visitor conditions. The system achieved an overall accuracy of 89.39%. Based on the classification report, the Crowded class achieved the best performance with an F1-score of 0.94, followed by the Quiet class (F1-score 0.91), while the Moderate class obtained the lowest F1-score of 0.82 due to its position between the two other classes. The main limitations of the system include occlusion conditions, lighting variation, and threshold sensitivity. The results indicate that the pretrained approach without retraining saves time and computational resources for the training process, while still producing competitive accuracy for visitor density classification applications in real-world environments.